InternW0: A Foundational Physical World Model for Efficient Real-World Interactions

TL;DR - InternW0 is a physical world model that jointly predicts future visual dynamics and generates continuous robot controls using experts operating at different timescales. Its asynchronous design aims to make embodied AI more efficient and responsive in changing, contact-rich real-world environments.

• Uses a high-capacity video expert for longer-horizon prediction and a lightweight action expert for faster control updates.
• Reuses layerwise K/V context and routes new observations into it, avoiding full future regeneration after every action.
• Trains on roughly 7,200 hours of heterogeneous robot and egocentric data, including the 275-hour EgoLab dataset.
• Supports varied robot embodiments and tactile, force-aware manipulation, with evaluations covering scientific synthesis and quantitative pipetting workflows.